Future Generation Computer Systems 20 (2004) 145–155
A virtual server queueing network method for component based performance modelling of metacomputing Yong Woo Lee The University of Seoul, The Faculty of Electrical and Computer Engineering, Seoul, South Korea
Abstract In this paper, we introduce a virtual server queueing network concept to evaluate the performance of metacomputing systems. It adds a virtual server concept to the existing queueing network theory. It enables us to effectively and accurately evaluate the performance of metacomputing systems. Component based software development now gains much interest and is especially regarded as very useful and necessary in the area of metacomputing. We find that the virtual server queueing network concept is well suited for a component based performance modelling and evaluation approach. In this paper, we have derived component based performance models of a metacomputing environment using the virtual server queueing network concept and validated them through measurement. With it, we quantitatively characterize the file access performance of caching in the metacomputing environment. We used a stochastic discrete event simulation as the main methodology to solve the models. © 2003 Elsevier B.V. All rights reserved. Keywords: Virtual server; Performance modelling; CBSD; Metacomputing; Caching
1. Introduction A lot of papers have been presented to study the performance of cache memory and related problems. Baker et al. [3] show the measured data for the file caches in the Sprite distributed system, discuss issues of file caching and show simulation results of a cache consistency mechanism. Ousterhout et al. [33] show the simulation results of file caching in local UNIX systems and discuss the issues of file caching such as file cache size, block size and write policy. Lilja [30] surveys cache coherence mechanisms in shared memory systems, discusses design issues, and studies the performance effect of the issues using trace driven simulations. Karedla et al. [23] discusses caching strategies and studies the E-mail address:
[email protected] (Y.W. Lee).
performance effect of cache replacement algorithms by simulation. Smith [43,44] discusses various cache memories and caching mechanisms in general and in detail. Also a lot of queueing network based papers have been presented to study the file access performance of distributed file systems. Ferrari et al. [15] show a model for VAX 11/780 systems. Bester et al. [6], Perros and Mirchandani [34], Lazowsak et al. [26] and Ramakrishnan and Emer [35,36] propose file access performance models of distributed file systems. However, we have had difficulties in evaluating the performance of complex computing systems such as metacomputing systems with queueing models [24,25]. In contrast to all these papers, in this paper, a virtual server concept is introduced. It adds a virtual server concept to the existing queueing network theory. It
0167-739X/$ – see front matter © 2003 Elsevier B.V. All rights reserved. doi:10.1016/S0167-739X(03)00141-9
146
Y.W. Lee / Future Generation Computer Systems 20 (2004) 145–155
enables us to use component based performance evaluation approach to investigate the file access performance of caching in metacomputing. Based on the operation of remote file access and local file access in each component, the two performance models were constructed. We measured real working environments and extracted the service time of the operation in each component. We validated the performance models and then projected the models further to investigate the file access performance of caching in metacomputing in which we are interested. This paper is organized as follows. Section 2 describes how we built our virtual server based performance models and explains our virtual server concept. We also explain our parameterization. Section 3 explains how we simulated our models. Section 4 describes the validation details of our models. In Section 5, we explore the file access performance of caching in metacomputing. Finally conclusion of this paper is given in Section 6.
2. Performance modelling In this section, we present our virtual server models to evaluate the performance of caching in metacomputing and describe our virtual server concepts. We will use component based performance modelling approach and see the inter-operation of components for caching in metacomputing.
2.1. Caching in local file access Fig. 1 shows the virtual server based performance model to evaluate the performance of caching in local file access. A local user sends a request for reading/writing a file/files. The command interpretation component interprets it first. It consumes the CPU power and is represented as a virtual server of the CPU. After interpretation, it receives two distinct services: the file handling by the file handling component and the disk I/O by the disk I/O component. The file handling operation consists of directory handling, file table lookup, updating file tables, opening files, closing files, etc. The file handling component does the file handling operation. The disk I/O operation consists of disk I/O path setup operation through the disk interface unit and physical disk I/O operation. The physical I/O operation consists of three major components: the seek, the latency and the transfer. The disk I/O component, together with the disk interface unit, does the disk I/O operation [1,2,3–5,7–9,14,16,18–20,22,23, 31,33,35,37,38,41,46–49]. If the request is a write request, the data are buffered to the memory first and then written into a disk. If it is a read request, the data are read first from the disk and then buffered to the memory. The file handling component consumes the CPU power and is represented as a virtual server of the CPU. The disk I/O component requires cooperation of both the CPU power and the disk. The required CPU power portion is served
I/O Local Users Memory
Cache
Disk-DMA Caching
CPU (C.I.)
CPU (File)
(Result Processing)
Cache
CPU (I/O)
CPU Disk DMA Disk
Fig. 1. Caching in local file access.
(Result Processing)
Y.W. Lee / Future Generation Computer Systems 20 (2004) 145–155
by a virtual server of the CPU. If cache hit occurs in the memory, all of the disk I/O related operations are bypassed with some cache processing overhead. If cache hit occurs in the disk interface unit (DMA), then pure disk I/O processing time is saved with some cache processing overhead. The read data are sent to the user screen or only the system message of reading/writing is sent to the user or no action is taken by the result processing component depending on the user request. This operation is served by a virtual server of the CPU and a screen display I/O server (result processing).
In the client, a user issues a request for reading/writing a file or files. The command interpretation component evaluates and interprets the request. Then file search operation is done by a virtual server of the CPU in order to check whether the request requires remote file access. If the requested file is in the local system, then it is processed as local file access. Otherwise, the RPC request building component makes a request, using RPC, of reading/writing the remote file from/to the designated file server. It is served by a virtual server of the CPU. The RPC request sending component, in cooperation with the network interface unit, sends the RPC request message via LAN to the file server which contains the requested file. In this operation, there can be certain degree of concurrency between the network interface unit and the CPU of the client [13,39,43]. It is done by a virtual server of the CPU and by a virtual server of the DMA [6,7,10,12,17,21,26,32,35, 40,45].
2.2. Caching in remote file access Fig. 2 shows the virtual server model to evaluate the performance of caching in remote file access. During simulation, it was assumed that, the file service activity in each client was so low that the contention for the system resources in the client was ignorable. CPU Client Caching(Disk)
147
Disk & Disk-DMA Memory
Client Caching(Memory)
Clients DMA CPU CPU (Net) (S) (B)
CPU (File)
CPU (C.I.)
I/O CPU (Result) (Result)
CPU (E)
CPU (R)
DMA (Net)
Retransmission Memory File Server Memory Caching
The File Server
Disk-DMA Caching I/O
DMA (Net)
CPU (S)
CPU (B)
Disk
Disk-DMA
CPU (I/O)
CPU (File)
Requeue Network
Fig. 2. Caching in remote file access.
CPU (E)
CPU (R)
DMA (Net)
148
Y.W. Lee / Future Generation Computer Systems 20 (2004) 145–155
After the network communication connection is established between the client and the designated file server, the request message traverses the LAN and arrives at the file server. There can be an operational delay between sending each request message from the client. This delay is called inter-request delay and depends on the characteristics of each metacomputing system. Retransmission can occur in case of unsuccessful transmission. The receiving operation at the file server is performed by RPC request receiving component of the file server. The RPC request sending component, the network and the network interface unit of the file sever cooperate to setup the communication connection to the file server and transfer the request message to the file server via LAN. Therefore, we specify a virtual server of the DMA and a virtual server of the CPU for it. The received request is stored in the buffers of the network interface unit. The buffered request message is sent to the memory of the file server for processing. The RPC request evaluation component fetches the request message and evaluates it. It is represented by a virtual server of the CPU. Once evaluated, the request is processed in the same way as a local request is processed in the local system. Now the RPC response building component makes a response message in response to the request message. A virtual server of the CPU in the model represents this operation. The response message is transferred from the memory to the network interface unit for sending. If the finite number of buffers of the network interface unit is already fully occupied, then the file server CPU should wait until the required buffer space is available. This is called requeue delay. The RPC response sending component, the network interface unit of file server, network, the RPC response receiving component of the client, and the network interface unit of the client cooperate to setup the communication connection to the client and transfer the queued response message to the client via LAN. The response message departs from the file server, traverses LAN and arrives at the client. In the model, it is served by a virtual server of the CPU in the file server, a virtual server of the DMA in the file server, the server of Network, a virtual server of the DMA
in the client and a virtual server of the CPU in the client. In Fig. 2, the CPU of the file server was represented by six virtual servers: the RPC request receiving virtual server, the RPC request evaluation virtual server, the file handling virtual server, the disk I/O virtual server, the RPC response building virtual server and the RPC response sending virtual server. Each virtual server competes for the real server as processes competes for the CPU in a computer system. The six virtual servers were mapped into the real server (the CPU) with a queue during our simulation. In our simulation, a real server is assigned to a virtual server when it is requested by a virtual server and the other virtual servers should wait to acquire the real server until the virtual server which is running releases it and it becomes free. Again in the client, the RPC response receiving component receives the response message in its finite buffers. The received response message is moved to the memory for processing. The RPC response evaluation component fetches the received response message and evaluates it. It is represented as a virtual server of the CPU. Now finally, the pure data processed by the result processing component are sent to the user’s screen of the client. This operation is served by a virtual server of the CPU and a screen display I/O server. The user using the client will repeat the above whole life cycle again or do thinking (it is often called as either the user think time or the idle time) or do stand-alone processing (it is often called local processing) in the client. 2.3. Parameterization Through specially designed measurement in several metacomputing environments, we quantified the parameters in Figs. 1 and 2. Tables 1 and 2 show the parameter values of the metacomputing environment which consisted of SUN SPARCstation 10 workstations. About 10 Mbps Ethernet was used.
3. Simulation We used stochastic discrete event simulation as the primary method to obtain the solution of our performance models. Simulation packages for queueing
Y.W. Lee / Future Generation Computer Systems 20 (2004) 145–155 Table 1 The service time of each operation by components in local file accessa Operation by components CPU CPU CPU I/O I/O CPU I/O
Command interpretation File handling Disk I/O Disk interface unit Disk interface unit + disk I/O Result processing Result processing
Execution time (ms) F F P F P P P
20.000 5.000 0.125 24.000 1.125 0.250 22.000
a CPU means that it is used as the CPU time and I/O means the I/O time. P means it is proportional to the data size and F means it is fixed (constant). All service times, proportional to the data size, are per 1500 bytes data transferred. The values of all service times, constant to the data size, are per one transaction regardless of the transferred data size.
network systems are GIST, NETWORK-II.5, NUMAS, PAWS, PANACEA, QNAP, RESQ, and RESQME. In our work, we used SLAM-II general simulation language. Fig. 3 shows a part of a sample implementation of our virtual server model concept Table 2 The service time of each operation by components in remote file accessa Operation by components CPU CPU CPU I/O I/O I/O CPU CPU CPU CPU I/O I/O CPU CPU I/O I/O I/O CPU CPU CPU I/O a
C C C C S S S S S S S S S S C C C C C
Command interpretation RPC request building RPC request sending Network interface unit Network transmission Network interface unit RPC request receiving RPC request evaluation File handling Disk I/O Disk interface unit Disk interface unit + disk I/O RPC response building RPC response sending Network interface unit Network transmission Network interface unit RPC response receiving RPC response evaluation Result processing Result processing
Execution time (ms) F F P P P P P F F P F P F P P P P P F P P
20.0000 1.2500 0.1125 0.2875 1.2000 0.2875 0.1125 1.2600 5.0000 0.1250 24.0000 1.1250 1.2500 0.1125 0.2875 1.2000 0.2875 0.1125 1.2500 0.2500 22.0000
S means it is used in the server and C means in the client.
149
...... INTLC,XX(29)=1.25 ; Constant : CPU Request Evaluation INTLC,XX(30)=5 ; Constant : CPU Request Processing ; (File Processing) ...... ;============================= NETWORK; ...... RESOURCE/4,SCPUK(1),4 ; Multiple CPU Numbers ...... ; Resource ID #, RName(# of Resources), ; Queue file number used in AWAIT ;============================= ...... ;============================= ; File server: Request Evaluation: CPU AWAIT(4),SCPUK; ACT/44,XX(29); FREE,SCPUK; ;============================= ; File server : Request Processing ; (File processing) : CPU ;============================= AWAIT(4),SCPUK; ACT/46,XX(30); FREE,SCPUK; ;============================= ......
Fig. 3. A SLAM-II implementation of the virtual server model of Fig. 2.
in SLAM-II for explanation. In Fig. 3, a real server is represented as a resource. “4”: “RESOURCE/4”. “SCPUK(1), 4” means it has one resource named as SCPUK and a queue with identification number “4’ is assigned for the resource. The queue “4” is specified as an FCFS queue but any queueing disciplines can be specified.
4. Measurement and validation We repeatedly performed various measurements specially designed to validate our models. Here we show a result for local file access in Table 3 and a result for remote file access in Table 4. Sun SPARCstation 10 workstations and 10 Mpbs Ethernet were
150
Y.W. Lee / Future Generation Computer Systems 20 (2004) 145–155
Table 3 A local read and display case Workload (Kbytes)
Measurement
Simulation
System time (ms)
1.5 15 150 300
Min
Max
70 100 720 1530
820 830 2830 5870
CPU time (ms) Mode 80 2730 5700
Min
Max
Mode
20 30 70 70
50 50 180 180
30
150
System time (ms)
CPU time (ms)
82 384 2756 5619
31 36 90 155
Table 4 A remote read and local display case Workload (Kbytes)
Measurement
Simulation
System time (ms)
1.5 15 150 300
Min
Max
100 130 320 530
170 220 4480 5020
CPU time (ms) Mode
Min
Max
Mode
50 80 100 170
40
1400
20 30 50 100
used for the experiments. In our experiments for both tables, we measured both system time and CPU time when there was no caching. Caching overhead was measured in other separate experiments. In the measurement for Table 3, a local read and display experiment, using “cat local file” command, was done. It read the ‘local file’ in the local disk and displayed the read data on the user screen. In the measurement for Table 4, a remote read and display experiment, using “cat remote file” command, was done. It read the ‘remote file’ in the disk of the file server and displayed the read data on the user screen. In the two tables, we specified the mode value, the most frequently observed value, if the frequency of the mode value was found to be more than 20% of the total occurrence. Otherwise, we left it blank. The accounting record showed the measured response time to the level of 10 ms.
130
System time (ms)
CPU time (ms)
126 147 709 1334
36 40 83 130
the Sun SPARCstation10 workstations and 10 Mbps Ethernet. Through the simulation of our component based performance models, we got statistics such as average (mean), maximum, minimum, standard deviation values of response time, queueing delay (waiting time) and system resource utilization. In this paper, we focus on the average response time among them as the index of the performance of caching. 5.1. Used workloads Through an elaborated workload characterization from real measured workloads in several environments, the workloads in Table 5 were obtained [3,11, 26–29,33–35,42]. The workload pattern of the Poisson distribution for input arrival and the log-normal distribution for input transaction size were used. 5.2. The stand-alone caching
5. Performance of caching This paper evaluates the performance of caching in the metacomputing environment which consisted of
In this sub-section, we investigate the performance of three kinds of cachings in remote file access. They are caching in the memory of the file server, caching
Y.W. Lee / Future Generation Computer Systems 20 (2004) 145–155
151
Table 5 The used workloads Case
1 2 3 4 5 6
W/L (Kbytes)
8 47 50.7 316 316 1856
Transaction size (Kbits per transaction)
Transaction number when the number of active clients = 100 (transactions/s)
Average
Standard deviation
Average
Standard deviation
64 376 405.6 2528 2528 14852
288 2144 768 6464 6464 37976
22.75 4.0 22.75 4.0 22.75 4.0
12.75 3.75 12.75 3.75 12.75 3.75
in the disk interface unit of the file server and caching in the client (local) memory. We also investigate the performance of two kinds of cachings in local file access. They are caching in the memory and caching in the disk interface unit. Here, stand-alone caching means that only one kind of caching is used in each file access. Fig. 4 presents the behaviour of the average response time in the five kinds of stand-alone cachings when the 50.7 Kbytes workload was used. The three workload cases in Fig. 5 show the effect of the caching of remote file access in the memory of the file server (at 0 (normal), 20, 40, 60, 80, 100% cache hit ratio, respectively).
(6.4%), respectively. In Fig. 8, we show a sample file access performance in the multiple levels of caches in different scenario. It shows the performance when the hit ratio in the disk interface unit of the file server were remote s-dma l-dma
local c-mem
s-mem l-mem
Avreage response time (msec)
1200
1000
5.3. Combination
800
Fig. 6 shows the file access performance in average response time when we used multiple levels of cachings at the same time by combining the stand-alone cachings. There, the following abbreviation is used. SA1 = s-mem. SA2 = c-dma. SA3 = c-mem. LS-SA1 = l-mem. LS-SA2 = l-dma. LS-CB: the combination of l-mem and l-dma; CB1: the combination of s-mem & c-mem. CB2: the combination of s-mem, s-dma & c-mem. CB3: the combination of s-mem & s-dma. In Fig. 7, we show sample file access performance of CB3 when the cache hit ratio in both caches were varied at the same rate. There, 60% cache hit at each cache in a file access (we call, relative cache hit ratio) means 60% at the first cache, 24% (=60% × (100% − 60%)) at the second cache and 12.8% at the third cache in actual hit ratio (we call, absolute hit ratio). Therefore in the case, the cumulative cache hit ratio (the cumulative miss ratio) is 60% (40%), 84% (16%) and 93.6%
600
400
200
0 0
10
20
30
40
50
60
Number of clients or local users Fig. 4. File access performance of stand-alone cachings. Abbreviation: “normal”—remote file access at 0% cache hit, “local”—local file access at 0% cache hit, “s-mem”—remote file access at 40% cache hit in the memory of the file server, “s-dma”—remote file access at 40% cache hit in the disk interface unit of the file server, “c-mem”—remote file access at 40% cache hit in the client memory, “l-mem”—local file access at 40% cache hit in the local memory, “l-dma”—local file access at 40% cache hit in the disk interface unit of the local system.
152
Y.W. Lee / Future Generation Computer Systems 20 (2004) 145–155
normal
20%
40%
normal
20%
40%
60%
80%
100%
60%
80%
100%
Average response time (m sec)
Average response time (sec)
300
7
250
6
200
5
150
4 3
100
2
50 1
0
0 (a)
20
40
60
80
100
0
0
Number of clients
50
(b)
normal
20%
60%
100%
100 150 200 Number of clients
250
300
40%
Average response time (sec) 14 12 10 8 6 4 2 0 0 (c)
5
10
Number of clients
Fig. 5. (a) The case of the 8 Kbytes workload; (b) the case of the 47 Kbytes workload; (c) the case of the 316 Kbytes workload.
Y.W. Lee / Future Generation Computer Systems 20 (2004) 145–155 SA 1 CB 1 LS-SA1
500
SA 2 CB 2 LS-SA2
SA 3 CB 3 LS-C B
1800
Average response time (m sec)
450
153
normal
20%
40%
60%
80%
100%
Average response time (m sec)
1600
400
1400
350 300
1200
250
1000
200 150
800
100 50
600
0 0
10 20 30 40 50 Number of clients or local users
400 200
Fig. 6. The average response time of the 50.7 Kbytes workload with 60% cache hit in each cache of various caching cases.
0 0
10 20 30 40 50 60 Number of clients or local users
Fig. 8. The effect on the average response time of the 50.7 Kbytes workload when the hit ratio varies only in the disk interface unit.
normal
20%
40%
60%
80%
100%
Average response time (m sec) 1800
varied, while the hit ratio in the memory of the file server were fixed at 60% in CB3. In Fig. 7, the cache hit ratio in both caches were varied simultaneously.
1600 1400
6. Conclusions
1200 1000 800 600 400 200 0 0 10 20 30 40 50 Number of clients or local users Fig. 7. The effect on the average response time of the 50.7 Kbytes workload when the hit ratio varies in the two levels of caches simultaneously.
In this paper, we proposed component based file access performance models to evaluate caching in metacomputing using our virtual server concept. In the models, each real server could be represented several virtual servers based on the operation in each component. The virtual server concept is based on queueing theory. Thus we could cheaply construct the performance models. Then through simulation we relatively easily obtained required performance values. Meanwhile, the obtained values showed relatively high accuracy compared with those of queueing models with analytic solutions. Also, we could easily modify our developed models to reflect another requirement. We could easily
154
Y.W. Lee / Future Generation Computer Systems 20 (2004) 145–155
extend our models to many modified case studies. We think it is because our virtual server concept is based on queueing theory and enables us to use component based performance modelling approach. In this study, we validated our models through measurement and used them to evaluate some caching issues. We studied both remote file access and local file access since both can occur in random manners and a subsystem can be either a client or a file server in real metacomputing environments.
Acknowledgements This study has been partially supported by the funds from KISTI, MIC (Ministry of Information and Communication) of Republic of Korea, SMBA (Small and Medium Business Administration) of Republic of Korea, Metatech Inc. and Ncerti Inc. References [1] J. Abate, H. Dubner, S. Weinberg, Queueing analysis of the IBM 2314 disk storage facility, J. ACM 15 (4) (1968) 577– 589. [2] H.P. Artis, DASD subsystems: evaluating the performance envelope, CMG Trans. (1994) 3–12. [3] M. Baker, J. Hartman, M. Kupfer, K. Shirriff, Measurements of a distributed file system, presented in the Proceedings of the 13th ACM Symposium on Operating System Principles, vol. 25, no. 5, October 1991, Oper. Syst. Rev. 198–212. [4] Y. Bard, A model of shared DASD and multipathing, CACM 23 (10) (1980). [5] S.J. Baylor, C. Benveniste, Y. Hsu, Performance evaluation of a massively parallel i/o subsystem, in: Proceedings of the IPPS’94, April 1994, Mexico, pp. 5–10. [6] J. Bester, et al., A dual priority MVA model for a large distributed system: LOCUS, in: Proceedings of the Performance’84, 1984. [7] L.N. Bhuyan, D. Ghosal, Q. Yang, Approximate analysis of single and multiple ring networks, IEEE Trans. Comput. 38 (7) (1989) 1027–1040. [8] A. Brandwajn, Models of DASD subsystems: basic model of reconnection, Perform. Eval. 1 (1981) 263–281. [9] A. Brandwajn, Models of DASD subsystems with multiple access paths: a throughput-driven approach, IEEE Trans. Comput. C-32 (5) (1983) 451–463. [10] W. Bux, Token ring local area networks and their performance, Proc. IEEE 77 (2) (1989) 238–256. [11] M. Calzarossa, D. Ferrari, A sensitivity study of the clustering approach to workload modeling, Perform. Eval. 6 (1986) 25– 33.
[12] D.D. Clark, et al., An analysis of TCP processing overhead, IEEE Commun. 27 (6) (1989) 23–29. [13] E. Cooper, et al., Protocol implementation on the nectar communication processor, in: Proceedings of the SIGCOMM 90 Symposium, Communication Architectures and Protocols, ACM Press, New York, 1990, pp. 135–143. [14] J.N. Daigle, R.B. Kuehl, J.D. Langford, Queueing analysis of an optical disk jukebox based office system, IEEE Trans. Comput. 39 (6) (1990). [15] D. Ferrari, et al., Modeling file system organizations in a local area network environment, Report No. UCB/CSD, 83/142, Progress Report No. 83.7, EECS, University of California, Berkeley, October 1983. [16] R.M. Geist, K.S. Trivedi, Optimal design of multilevel storage hierarchies, IEEE Trans. Comput. C-31 (3) (1982) 249–260. [17] A. Goldberg, G. Popek, S.S. Levenberg, A validated distributed system performance model, in: A.K. Agrawala, S.K. Tripathi (Eds.), Proceedings of the PERFORMANCE’83, North-Holland, Amsterdam, The Netherlands, 1983, pp. 251–268. [18] C.C. Gotlieb, G.H. Macewen, Performance of moveable-head disk storage devices, J. ACM 20 (4) (1973) 604–623. [19] A. Goyal, T. Agerwala, Performance analysis of future shared storage systems, IBM J. Res. Dev. 28 (1) (1984) 95–108. [20] G.E. Houtekamer, The local disk controller, ACM SIGMETRICS Perform. Eval. Rev. 13 (2) (1985) (special issue); G.E. Houtekamer, Proceedings of the 1985 ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems, August 1985. [21] R. Jain, Performance analysis of FDDI token ring networks: effect of parameters and guidelines for setting TTRT, in: Proceedings of the SIGCOMM’90 Symposium: Communication Architectures and Protocols, September 24–27, 1990, Comput. Commun. Rev. 20 (4) (1990). [22] M.Y. Kim, Synchronized disk interleaving, IEEE Trans. Comput. 35 (11) (1986) 978–988. [23] R. Karedla, J.S. Love, B.G. Wherry, Caching strategies to improve disk system performance, IEEE Comput. (1994) 38–46. [24] S.S. Lavenberg, A perspective on queueing models of computer performance, Perform. Eval. 10 (1) (1989) 53–76. [25] S.S. Lavenberg, Performance evaluation in industry: a personal perspective, Perform. Eval. (2000) 3–13. [26] E.D. Lazowsak, J. Zahorjan, D.R. Cheriton, W. Zwaenepoel, File access performance of diskless workstations, ACM Trans. Comput. Syst. 4 (3) (1986). [27] Y.W. Lee, A.S. Wight, D.H. Lee, Performance modeling and simulation of the 1993 Daejeon International EXPO network, in: Proceedings of the UK Simulation Society’93 Conference, Keswick, UK, September 1993. [28] Y.W. Lee, A.S. Wight, Y.W. CHO, Workload characterization of Cray supercomputer systems running UNICOS for the optimal design of NQS configuration in a site, in: Proceedings of the 33rd International Cray User Group Conference, San Diego, CA, USA, March 14–18, 1994. [29] Y.W. Lee, A.S. Wight, S.W. Choi, D.H. Lee, Simulation of compound local area networks for a large scale client server
Y.W. Lee / Future Generation Computer Systems 20 (2004) 145–155
[30]
[31] [32]
[33]
[34]
[35]
[36]
[37]
[38] [39] [40] [41] [42]
type multi-media computer system, in: Proceedings of the International Pritsker User Group Conference, June 7–9, 1995, Indianapolis, USA. D.J. Lilja, Cache coherence in large-scale shared memory multiprocessors: issues and comparisons, ACM Comput. Surv. 25 (3) (1993) 303–338. J.B. Major, Processor, I/O path, and DASD configuration capacity, IBM Syst. J. 20 (1) (1981). M. Marathe, S. Kumar, Analytic models for an Ethernet-like LAN link, in: Proceedings of the ACM SIGMETRICS Conference Proceedings, September 1981. J.K. Ousterhout, H.D. Costa, D. Harrison, J.A. Kunze, M. Kupfer, J.G. Thompson, A trace driven analysis of the UNIX 4.2 BSD system, in: Proceedings of the 10th ACM Symposium on Operating System, December 1–4, 1985. H.G. Perros, D. Mirchandani, An analytic model of a file server for bulk file transfer, ACM SIGMETRICS Perform. Eval. Rev. 1.3 (3–4) (1985) 14–22. K.K. Ramakrishnan, J.S. Emer, A model of file server performance for a heterogeneous distributed system, in: Proceedings of the SIGCOMM’86 Symposium: Communications Architectures and Protocols, August 1986, pp. 338–347. K.K. Ramakrishnan, J.S. Emer, Performance analysis of mass storage service alternatives for distributed systems, IEEE Trans. Softw. Eng. 15 (2) (1989). C. Rhuemmler, J. Wilkes, UNIX disk access patterns, in: Proceedings of the 1993 Winter USENIX, January 25–29, 1993, San Diego, USA. C. Rhuemmler, J. Wilkes, An introduction to disk drive modeling, IEEE Comput. (1994) 17–28. M. Schroeder, M. Burrows, Performance of firefly RPC, ACM Trans. Comput. Syst. 8 (1) (1990) 1–17. J.F. Shoch, J.A. Hupp, Measured performance of an Ethernet local network, CACM 23 (12) (1980) 711–721. C.E. Skinner, J.R. Asher, Effects of storage connection on system performance, IBM Syst. J. 4 (1969) 319–333. A.J. Smith, Long term file migration: development and evaluation of algorithms, CACM 24 (8) (1981) 521–532.
155
[43] A.J. Smith, Cache memories, ACM Comput. Surv. 14 (3) (1982) 473–530. [44] A.J. Smith, Disk-cache miss-ratio analysis and design considerations, ACM Trans. Comput. Syst. 3 (3) (1985) 161–203. [45] P.A. Steenkiste, A systematic approach to host interface design for high speed networks, IEEE Comput. (1994) 47–57. [46] Teorey, Pinkerton, A comparative analysis of disk scheduling policies, CACM 15 (3) (1972) 177–184. [47] S.J. Waters, Estimating magnetic disk seeks, Comput. J. 18 (1975) 12–19. [48] N. Wilhelm, A general model for the performance of disk systems, J. ACM 240 (1) (1977) 14–31. [49] C. Wood, P. Hodges, DASD trends: cost, performance, and form factor, Proc. IEEE 81 (4) (1993) 573–585.
Yong Woo Lee received the BS degree in electrical engineering from Seoul National University, Seoul, South Korea, in 1981 and PhD degree in computer science from The University of Edinburgh, Edinburgh, Scotland, Great Britain in 1997. He is currently the director of Seoul GRID Center and an associate professor in the division of Electrical, Electronics and Computer Engineering (EECE) at The University of Seoul, Seoul, South Korea. He was a principal research scientist at The National Multimedia Education Center of the Ministry of Education, Seoul, South Korea in 1998 and a senior research scientist at Korea Institute of Science and Technology (KIST) and Korea Advanced Institute of Science and Technology (KAIST) from 1982 to 1998. He was also an international engineer at Schlumberger Technical Service Inc. in 1981. His research interests are in GRID (metacomputing), CSCW, component based system software, performance evaluation engineering, computer communication, content delivery and distribution, Internet, high performance computing and IT in education.